Canaan Creative's AI chip business

Source/Leifeng.com

Will mining machine giants become AI chip giants in the future?

Author | Bao Yonggang

Produced by | Leifeng.com Industry Group

Canaan Inc., the world's first blockchain stock, is about to celebrate its first anniversary since its listing. Looking back at its performance over the past year, its stock price fluctuated with the price of Bitcoin.

In recent months, the rise in Bitcoin prices has made Canaan Technology's Avalon mining machines hard to come by, with orders already scheduled until March next year.

Compared with the uncertainty of Bitcoin prices, the rapid growth of AI chips in the next few years is very clear, and growth is accompanied by opportunities. In 2018, many AI chips competed with each other in performance through PPT, and Canaan Technology launched its first AI chip, which is also the world's earliest mass-produced RISC-V architecture commercial edge AI chip.

In 2020, a key year for the implementation of AI chips, Canaan Technology expects its AI products to bring in 10 million high-quality revenue, and also expects AI chip revenue to reach 100 million yuan in the next 2-3 years. Why can the AI ​​chip of Canaan Technology, the world's second largest mining machine company, be implemented so quickly? Will the mining machine giant become an AI chip giant in the future?

1

Design AI chips with the ability to design mining chips

Walking on two legs is definitely better than walking on one leg. This can be used to explain why Canaan Technology continued to expand the market after its success in the blockchain market, but why did it choose AI? In 2013, when Canaan Technology was founded, the computing power of ASIC (application-specific integrated circuit) in Bitcoin mining exceeded everyone's imagination. With the timely launch of the Avalon mining machine and continuous iteration, Canaan Technology has achieved success in the blockchain market, which has strict requirements on computing power and power consumption.

"2016 was a very important milestone for us. We completed all the work on the most advanced 16nm process at the time, and then began to look for the next technology and product that could give full play to our chip design capabilities." Zhang Nangeng, founder, chairman and CEO of Canaan Technology, said in a conversation with Leifeng.com, "There are two main reasons why we finally chose AI. One is that AI algorithms do have a strong demand for repeated calculations, and the other is that ASIC, which we are good at, does have a place to use."

Zhang Nangeng, Founder, Chairman and CEO of Canaan Inc.

In 2016, a new round of AI craze emerged. AI algorithms were rapidly iterating, and the computing power of traditional chips was far from meeting the demand. AI-specific chips could significantly improve the results, which triggered a wave of AI chip entrepreneurship. Canaan Technology, which has accumulated experience in the chip field, also saw this opportunity and chose the edge AI market.

The cloud AI chip market is lucrative, but large companies such as Nvidia and Intel have an advantage. The edge market is not very profitable but has huge potential. Why did Canaan Technology, which is good at designing high-performance dedicated chips, choose the edge AI chip market?

Zhang Nangeng said: "Cloud AI chips have some trade-offs in order to achieve ideal performance. They belong to the more traditional chip design field. This market is crowded with giants, leaving few opportunities for start-ups. There is no leading company on the edge, and both edge AI and mining machine chips have extreme requirements on cost, power consumption and performance. The advantages of our chip design can be well utilized on edge AI chips."

Somewhat surprisingly, Canaan Technology decided to use the RISC-V instruction set when it launched its AI chip project in 2016. You should know that the RISC-V project was launched at the University of California, Berkeley in 2010, and the first version of the user manual was officially released in 2014. It began to attract attention in China in 2018. Even now, many companies are still waiting to see whether to use RISC-V.

"Indecision is actually a waste of time. There are many geeks in our team, and we think RISC-V has more promise, so we decided on it," Zhang Nangeng explained.

2

The earliest mass-produced RISC-V edge AI chip,

Is the educational robot market “unrivaled”?

An important reason why Canaan did not waste time wavering was that it carefully analyzed its product positioning. Zhang Nangeng said: "We are not rejecting the Arm platform, but because many IoT devices do not have screens and do not run Android systems, the Arm ecosystem does not have such a big advantage. RISC-V can make good use of its customizability in the IoT market, and it is easier to expand."

Being the first to use RISC-V naturally faces greater challenges. "Although using RISC-V saves a lot of IP licensing fees, there are many pitfalls, so the R&D cycle will be longer. I think overall RISC-V does not save much cost."

Zhang Nangeng said, "It's not so bad for larger bugs. Sometimes, when a chip has been taped out and random errors occur, we can only use the compiler to solve them, or even cut some features. Of course, this is difficult to avoid for any complex chip or system."

Looking back at the choice made back then, Zhang Nangeng said that choosing RISC-V was a very appropriate choice. RISC-V is now in use, and the next step is to develop in the direction of customization.

The first product that Canaan used RISC-V in was the K210 released in 2018. This AI chip consumes less than 0.3W of power, less than 1W of power in typical application scenarios, and has an AI computing power of 1TOPS. Compared with the newly launched AI chips, the AI ​​computing power of 1TOPS is indeed not high, but this is closely related to its product positioning and launch time.

"When we defined the K210, we didn't have a deep understanding of AI models. MobileNetV2 and YOLOv3 algorithms were not available when we invested in the chip, so the requirement was very simple, which was 1TOPS of AI computing power." Zhang Nangeng further stated, "We also developed the AI ​​accelerator ourselves. Compared with AI accelerators of the same period, Canaan's accelerator is relatively more flexible."

With the evolution of AI technology, there are also application scenarios at the edge that have higher requirements for AI computing power. Zhang Nangeng revealed that Canaan's next-generation AI accelerator uses a scalable architecture that can achieve higher computing power. At the same time, versatility is further enhanced. In principle, models using TensorFlow do not need to be recompiled or trained.

At present, Canaan's first AI chip product has achieved good results. Leifeng.com learned that based on Kanzhi K210, Canaan has developed multiple product lines including intelligent touchless access control, intelligent attendance machine and intelligent meter reading, realizing the intelligentization of equipment in enterprise parks and intelligent meters for civil and commercial use. In the smart city, the largest scenario of AIoT application, Canaan has established cooperation with Haidian City Brain and Jianggan District of Hangzhou.

It is particularly worth mentioning that Canaan has won many leading customers in the educational robot market. For example, the uKit2.0 smart car project, a collaboration between Canaan and UBTECH in the field of AI STEAM education, has entered the trial production stage, which seems to be unrivaled.

Zhang Nangeng said: "Many of our customers in the education market are very well-known. They evaluated for a long time, but finally chose us. The reason is that in addition to the fact that there are only a few end-side AI chips, our products have been launched for some time, have better support for models, and are easier to develop educational courses. In addition, the price of around US$3 is also very attractive."

With the large-scale deployment of the first generation of AI chip products, Canaan expects to achieve 10 million high-quality AI revenue in 2020. Regarding this expectation, Zhang Nangeng said: "I am still very confident about the 10 million AI revenue forecast this year, and this is a relatively high-quality revenue, because we do not do system integration, this revenue is very pure chips and modules, plus a small amount of algorithm revenue."

3

Canaan AI's revenue is expected to reach 100 million yuan in the next 3-5 years

Revenues of tens of millions are not much for a company like Canaan with a market value of over US$500 million, but it is a good result in the early AI edge chip market.

"A chip generally needs a market cultivation period of 1-2 years before it can be rapidly mass-produced, and the cultivation period for AI chips is even longer. In particular, our chip CPU uses RISC-V, which requires some time to get started. In addition, there are very few people who understand AI, and edge AI chips are still a bit far from an explosion." Zhang Nangeng pointed out.

It can be seen that Canaan has also made some achievements in the construction of the ecosystem. In April this year, Canaan K210 obtained support for the mainstream embedded operating system Linux kernel. At the same time, the system-level language RUST can also be supported.

In addition, Canaan has established cooperation with Baidu PaddlePaddle, the largest algorithm platform in China, to develop and customize the edge AI development board PaddlePi-K210, which can provide developers with a one-stop development experience. Kanzhi K210 has also attracted the attention of Alibaba and is an exploratory learning platform in the direction of TinyML.

Canaan Technology AI chip ecosystem partner

The competitiveness of chip companies is ultimately due to the advantages of the ecosystem. The chips themselves are highly replaceable, but it is much more difficult to replace both hardware and software at the same time. Unlike many successful chip companies that have built a closed ecosystem, Canaan emphasizes openness when building an ecosystem. Zhang Nangeng explained: "We are an open source culture in blockchain, and RISC-V is also open source. I am now leading the effort to make more of the underlying things of the K210 open source. We will pay more attention to open source from the beginning of the next generation of products."

Canaan's openness is also reflected in its products. After the release of their AI chips, it is easy to buy the corresponding modules and development boards. Of course, this is inseparable from Canaan's experience in the chip industry chain. The openness of hardware from technology to sales, and the open source of software from documents to systems have allowed Canaan to expand its ecosystem step by step.

Zhang Nangeng predicts that based on the rapid growth trend of its current customer base and sales, Canaan's AI revenue will increase by an order of magnitude in the next 3-5 years, reaching the level of 100 million yuan.

Data from market consulting firm ABI Research also shows that by 2025, edge AI chip market revenue is expected to reach US$12.2 billion, and cloud AI chip revenue is expected to reach US$11.9 billion. The edge AI chip market size will exceed cloud AI chip market size for the first time.

Canaan's billion-yuan AI revenue will come from replacement and incremental markets. The so-called replacement market is to upgrade and replace existing chips with smart ones, such as making door locks and surveillance cameras smarter. The incremental market is some new applications, such as face tracking for selfie gimbals, which realize some new functions that were previously difficult to achieve through lower costs and higher performance.

It is reported that Canaan's AI product line will be divided into two series based on performance, and will maintain an iteration speed of one generation every two years, exploring more application areas including smart retail, ADAS assisted driving and smart finance.

Regarding the changes that 5G+AI will bring to the AI ​​market, Zhang Nangeng believes that this is a big variable and it is difficult to predict where it will eventually go.

4

The absolute value of blockchain and AI revenues must continue to grow

Canaan has attracted the attention of the industry with its Avalon series of mining machines. The experience accumulated in designing mining machine ASIC chips has enabled Canaan to launch AI chips earlier when the AI ​​wave comes. Canaan's experience in the chip supply chain and open mind have helped their AI chips to be implemented faster and generate revenue, and the educational robot market has become a representative scenario for its implementation.

In addition, the edge AI chip market is not a lucrative market, but the space in this market is large enough. Being able to realize earlier that chips need to consider software implementation will also help Canaan to better expand into more markets.

As for whether Canaan's AI revenue will exceed that of blockchain in the future, Zhang Nangeng's attitude is that the two businesses are not in competition within the company, but he just hopes that the absolute value of the company's revenue will continue to grow.